A Privacy-Preserving Federated Architecture for Cross-Institutional Fraud-Signal Sharing Among Commercial Banks, Credit Unions, and Community Development Financial Institutions

Abstract

Cross-institutional fraud detection faces a persistent structural problem: financial crime routinely spans multiple banks, credit unions, and Community Development Financial Institutions (CDFIs), yet fewer than one in eight eligible U.S. financial institutions currently participate in FinCEN's Section 314(b) voluntary information-sharing programme, with roughly 7,200 institutions registered against a base that runs into the tens of thousands once every eligible entity type is counted. Federated learning has emerged as a technically credible alternative that allows institutions to collaboratively train fraud-detection models without centralising raw customer data, and a growing body of published research now reports concrete performance results rather than purely theoretical promise. This article reviews published, cited evidence on privacy-preserving federated architectures for cross-institutional fraud-signal sharing, including a peer-reviewed federated meta-learning framework for credit card fraud detection that outperformed ten state-of-the-art baseline models, and a published architecture for differentially private secure multi-party computation in federated financial-services applications. The article situates these results against FinCEN's Section 314(b) safe-harbour framework, which already covers fraud-specific information sharing even though the sharing process itself remains largely manual, reviews the architecture such systems require, and candidly assesses the evidentiary gaps that remain before this technology can be considered validated for community financial institutions specifically.